2 * Copyright (c) 1991 Regents of the University of California.
4 * Copyright (c) 1994 John S. Dyson
6 * Copyright (c) 1994 David Greenman
9 * This code is derived from software contributed to Berkeley by
10 * The Mach Operating System project at Carnegie-Mellon University.
12 * Redistribution and use in source and binary forms, with or without
13 * modification, are permitted provided that the following conditions
15 * 1. Redistributions of source code must retain the above copyright
16 * notice, this list of conditions and the following disclaimer.
17 * 2. Redistributions in binary form must reproduce the above copyright
18 * notice, this list of conditions and the following disclaimer in the
19 * documentation and/or other materials provided with the distribution.
20 * 3. All advertising materials mentioning features or use of this software
21 * must display the following acknowledgement:
22 * This product includes software developed by the University of
23 * California, Berkeley and its contributors.
24 * 4. Neither the name of the University nor the names of its contributors
25 * may be used to endorse or promote products derived from this software
26 * without specific prior written permission.
28 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
29 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
30 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
31 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
32 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
33 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
34 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
35 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
36 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
37 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
40 * from: @(#)vm_pageout.c 7.4 (Berkeley) 5/7/91
43 * Copyright (c) 1987, 1990 Carnegie-Mellon University.
44 * All rights reserved.
46 * Authors: Avadis Tevanian, Jr., Michael Wayne Young
48 * Permission to use, copy, modify and distribute this software and
49 * its documentation is hereby granted, provided that both the copyright
50 * notice and this permission notice appear in all copies of the
51 * software, derivative works or modified versions, and any portions
52 * thereof, and that both notices appear in supporting documentation.
54 * CARNEGIE MELLON ALLOWS FREE USE OF THIS SOFTWARE IN ITS "AS IS"
55 * CONDITION. CARNEGIE MELLON DISCLAIMS ANY LIABILITY OF ANY KIND
56 * FOR ANY DAMAGES WHATSOEVER RESULTING FROM THE USE OF THIS SOFTWARE.
58 * Carnegie Mellon requests users of this software to return to
60 * Software Distribution Coordinator or Software.Distribution@CS.CMU.EDU
61 * School of Computer Science
62 * Carnegie Mellon University
63 * Pittsburgh PA 15213-3890
65 * any improvements or extensions that they make and grant Carnegie the
66 * rights to redistribute these changes.
68 * $FreeBSD: src/sys/vm/vm_pageout.c,v 1.151.2.15 2002/12/29 18:21:04 dillon Exp $
69 * $DragonFly: src/sys/vm/vm_pageout.c,v 1.26 2006/11/07 17:51:24 dillon Exp $
73 * The proverbial page-out daemon.
77 #include <sys/param.h>
78 #include <sys/systm.h>
79 #include <sys/kernel.h>
81 #include <sys/kthread.h>
82 #include <sys/resourcevar.h>
83 #include <sys/signalvar.h>
84 #include <sys/vnode.h>
85 #include <sys/vmmeter.h>
86 #include <sys/sysctl.h>
89 #include <vm/vm_param.h>
91 #include <vm/vm_object.h>
92 #include <vm/vm_page.h>
93 #include <vm/vm_map.h>
94 #include <vm/vm_pageout.h>
95 #include <vm/vm_pager.h>
96 #include <vm/swap_pager.h>
97 #include <vm/vm_extern.h>
99 #include <sys/thread2.h>
100 #include <vm/vm_page2.h>
103 * System initialization
106 /* the kernel process "vm_pageout"*/
107 static void vm_pageout (void);
108 static int vm_pageout_clean (vm_page_t);
109 static void vm_pageout_scan (int pass);
110 static int vm_pageout_free_page_calc (vm_size_t count);
111 struct thread *pagethread;
113 static struct kproc_desc page_kp = {
118 SYSINIT(pagedaemon, SI_SUB_KTHREAD_PAGE, SI_ORDER_FIRST, kproc_start, &page_kp)
120 #if !defined(NO_SWAPPING)
121 /* the kernel process "vm_daemon"*/
122 static void vm_daemon (void);
123 static struct thread *vmthread;
125 static struct kproc_desc vm_kp = {
130 SYSINIT(vmdaemon, SI_SUB_KTHREAD_VM, SI_ORDER_FIRST, kproc_start, &vm_kp)
134 int vm_pages_needed=0; /* Event on which pageout daemon sleeps */
135 int vm_pageout_deficit=0; /* Estimated number of pages deficit */
136 int vm_pageout_pages_needed=0; /* flag saying that the pageout daemon needs pages */
138 #if !defined(NO_SWAPPING)
139 static int vm_pageout_req_swapout; /* XXX */
140 static int vm_daemon_needed;
142 extern int vm_swap_size;
143 static int vm_max_launder = 32;
144 static int vm_pageout_stats_max=0, vm_pageout_stats_interval = 0;
145 static int vm_pageout_full_stats_interval = 0;
146 static int vm_pageout_stats_free_max=0, vm_pageout_algorithm=0;
147 static int defer_swap_pageouts=0;
148 static int disable_swap_pageouts=0;
150 #if defined(NO_SWAPPING)
151 static int vm_swap_enabled=0;
152 static int vm_swap_idle_enabled=0;
154 static int vm_swap_enabled=1;
155 static int vm_swap_idle_enabled=0;
158 SYSCTL_INT(_vm, VM_PAGEOUT_ALGORITHM, pageout_algorithm,
159 CTLFLAG_RW, &vm_pageout_algorithm, 0, "LRU page mgmt");
161 SYSCTL_INT(_vm, OID_AUTO, max_launder,
162 CTLFLAG_RW, &vm_max_launder, 0, "Limit dirty flushes in pageout");
164 SYSCTL_INT(_vm, OID_AUTO, pageout_stats_max,
165 CTLFLAG_RW, &vm_pageout_stats_max, 0, "Max pageout stats scan length");
167 SYSCTL_INT(_vm, OID_AUTO, pageout_full_stats_interval,
168 CTLFLAG_RW, &vm_pageout_full_stats_interval, 0, "Interval for full stats scan");
170 SYSCTL_INT(_vm, OID_AUTO, pageout_stats_interval,
171 CTLFLAG_RW, &vm_pageout_stats_interval, 0, "Interval for partial stats scan");
173 SYSCTL_INT(_vm, OID_AUTO, pageout_stats_free_max,
174 CTLFLAG_RW, &vm_pageout_stats_free_max, 0, "Not implemented");
176 #if defined(NO_SWAPPING)
177 SYSCTL_INT(_vm, VM_SWAPPING_ENABLED, swap_enabled,
178 CTLFLAG_RD, &vm_swap_enabled, 0, "");
179 SYSCTL_INT(_vm, OID_AUTO, swap_idle_enabled,
180 CTLFLAG_RD, &vm_swap_idle_enabled, 0, "");
182 SYSCTL_INT(_vm, VM_SWAPPING_ENABLED, swap_enabled,
183 CTLFLAG_RW, &vm_swap_enabled, 0, "Enable entire process swapout");
184 SYSCTL_INT(_vm, OID_AUTO, swap_idle_enabled,
185 CTLFLAG_RW, &vm_swap_idle_enabled, 0, "Allow swapout on idle criteria");
188 SYSCTL_INT(_vm, OID_AUTO, defer_swapspace_pageouts,
189 CTLFLAG_RW, &defer_swap_pageouts, 0, "Give preference to dirty pages in mem");
191 SYSCTL_INT(_vm, OID_AUTO, disable_swapspace_pageouts,
192 CTLFLAG_RW, &disable_swap_pageouts, 0, "Disallow swapout of dirty pages");
194 static int pageout_lock_miss;
195 SYSCTL_INT(_vm, OID_AUTO, pageout_lock_miss,
196 CTLFLAG_RD, &pageout_lock_miss, 0, "vget() lock misses during pageout");
199 SYSCTL_INT(_vm, OID_AUTO, vm_load,
200 CTLFLAG_RD, &vm_load, 0, "load on the VM system");
201 int vm_load_enable = 1;
202 SYSCTL_INT(_vm, OID_AUTO, vm_load_enable,
203 CTLFLAG_RW, &vm_load_enable, 0, "enable vm_load rate limiting");
206 SYSCTL_INT(_vm, OID_AUTO, vm_load_debug,
207 CTLFLAG_RW, &vm_load_debug, 0, "debug vm_load");
210 #define VM_PAGEOUT_PAGE_COUNT 16
211 int vm_pageout_page_count = VM_PAGEOUT_PAGE_COUNT;
213 int vm_page_max_wired; /* XXX max # of wired pages system-wide */
215 #if !defined(NO_SWAPPING)
216 typedef void freeer_fcn_t (vm_map_t, vm_object_t, vm_pindex_t, int);
217 static void vm_pageout_map_deactivate_pages (vm_map_t, vm_pindex_t);
218 static freeer_fcn_t vm_pageout_object_deactivate_pages;
219 static void vm_req_vmdaemon (void);
221 static void vm_pageout_page_stats(void);
227 vm_fault_ratecheck(void)
229 if (vm_pages_needed) {
241 * Clean the page and remove it from the laundry. The page must not be
244 * We set the busy bit to cause potential page faults on this page to
245 * block. Note the careful timing, however, the busy bit isn't set till
246 * late and we cannot do anything that will mess with the page.
250 vm_pageout_clean(vm_page_t m)
253 vm_page_t mc[2*vm_pageout_page_count];
255 int ib, is, page_base;
256 vm_pindex_t pindex = m->pindex;
261 * It doesn't cost us anything to pageout OBJT_DEFAULT or OBJT_SWAP
262 * with the new swapper, but we could have serious problems paging
263 * out other object types if there is insufficient memory.
265 * Unfortunately, checking free memory here is far too late, so the
266 * check has been moved up a procedural level.
270 * Don't mess with the page if it's busy, held, or special
272 if ((m->hold_count != 0) ||
273 ((m->busy != 0) || (m->flags & (PG_BUSY|PG_UNMANAGED)))) {
277 mc[vm_pageout_page_count] = m;
279 page_base = vm_pageout_page_count;
284 * Scan object for clusterable pages.
286 * We can cluster ONLY if: ->> the page is NOT
287 * clean, wired, busy, held, or mapped into a
288 * buffer, and one of the following:
289 * 1) The page is inactive, or a seldom used
292 * 2) we force the issue.
294 * During heavy mmap/modification loads the pageout
295 * daemon can really fragment the underlying file
296 * due to flushing pages out of order and not trying
297 * align the clusters (which leave sporatic out-of-order
298 * holes). To solve this problem we do the reverse scan
299 * first and attempt to align our cluster, then do a
300 * forward scan if room remains.
304 while (ib && pageout_count < vm_pageout_page_count) {
312 if ((p = vm_page_lookup(object, pindex - ib)) == NULL) {
316 if (((p->queue - p->pc) == PQ_CACHE) ||
317 (p->flags & (PG_BUSY|PG_UNMANAGED)) || p->busy) {
321 vm_page_test_dirty(p);
322 if ((p->dirty & p->valid) == 0 ||
323 p->queue != PQ_INACTIVE ||
324 p->wire_count != 0 || /* may be held by buf cache */
325 p->hold_count != 0) { /* may be undergoing I/O */
333 * alignment boundry, stop here and switch directions. Do
336 if ((pindex - (ib - 1)) % vm_pageout_page_count == 0)
340 while (pageout_count < vm_pageout_page_count &&
341 pindex + is < object->size) {
344 if ((p = vm_page_lookup(object, pindex + is)) == NULL)
346 if (((p->queue - p->pc) == PQ_CACHE) ||
347 (p->flags & (PG_BUSY|PG_UNMANAGED)) || p->busy) {
350 vm_page_test_dirty(p);
351 if ((p->dirty & p->valid) == 0 ||
352 p->queue != PQ_INACTIVE ||
353 p->wire_count != 0 || /* may be held by buf cache */
354 p->hold_count != 0) { /* may be undergoing I/O */
357 mc[page_base + pageout_count] = p;
363 * If we exhausted our forward scan, continue with the reverse scan
364 * when possible, even past a page boundry. This catches boundry
367 if (ib && pageout_count < vm_pageout_page_count)
371 * we allow reads during pageouts...
373 return vm_pageout_flush(&mc[page_base], pageout_count, 0);
377 * vm_pageout_flush() - launder the given pages
379 * The given pages are laundered. Note that we setup for the start of
380 * I/O ( i.e. busy the page ), mark it read-only, and bump the object
381 * reference count all in here rather then in the parent. If we want
382 * the parent to do more sophisticated things we may have to change
387 vm_pageout_flush(vm_page_t *mc, int count, int flags)
390 int pageout_status[count];
395 * Initiate I/O. Bump the vm_page_t->busy counter and
396 * mark the pages read-only.
398 * We do not have to fixup the clean/dirty bits here... we can
399 * allow the pager to do it after the I/O completes.
402 for (i = 0; i < count; i++) {
403 KASSERT(mc[i]->valid == VM_PAGE_BITS_ALL, ("vm_pageout_flush page %p index %d/%d: partially invalid page", mc[i], i, count));
404 vm_page_io_start(mc[i]);
405 vm_page_protect(mc[i], VM_PROT_READ);
408 object = mc[0]->object;
409 vm_object_pip_add(object, count);
411 vm_pager_put_pages(object, mc, count,
412 (flags | ((object == kernel_object) ? VM_PAGER_PUT_SYNC : 0)),
415 for (i = 0; i < count; i++) {
416 vm_page_t mt = mc[i];
418 switch (pageout_status[i]) {
427 * Page outside of range of object. Right now we
428 * essentially lose the changes by pretending it
431 pmap_clear_modify(mt);
437 * If page couldn't be paged out, then reactivate the
438 * page so it doesn't clog the inactive list. (We
439 * will try paging out it again later).
441 vm_page_activate(mt);
448 * If the operation is still going, leave the page busy to
449 * block all other accesses. Also, leave the paging in
450 * progress indicator set so that we don't attempt an object
453 if (pageout_status[i] != VM_PAGER_PEND) {
454 vm_object_pip_wakeup(object);
455 vm_page_io_finish(mt);
456 if (!vm_page_count_severe() || !vm_page_try_to_cache(mt))
457 vm_page_protect(mt, VM_PROT_READ);
463 #if !defined(NO_SWAPPING)
465 * vm_pageout_object_deactivate_pages
467 * deactivate enough pages to satisfy the inactive target
468 * requirements or if vm_page_proc_limit is set, then
469 * deactivate all of the pages in the object and its
472 * The object and map must be locked.
475 vm_pageout_object_deactivate_pages(vm_map_t map, vm_object_t object,
476 vm_pindex_t desired, int map_remove_only)
482 if (object->type == OBJT_DEVICE || object->type == OBJT_PHYS)
486 if (pmap_resident_count(vm_map_pmap(map)) <= desired)
488 if (object->paging_in_progress)
491 remove_mode = map_remove_only;
492 if (object->shadow_count > 1)
496 * scan the objects entire memory queue. spl protection is
497 * required to avoid an interrupt unbusy/free race against
501 rcount = object->resident_page_count;
502 p = TAILQ_FIRST(&object->memq);
504 while (p && (rcount-- > 0)) {
506 if (pmap_resident_count(vm_map_pmap(map)) <= desired) {
510 next = TAILQ_NEXT(p, listq);
511 mycpu->gd_cnt.v_pdpages++;
512 if (p->wire_count != 0 ||
513 p->hold_count != 0 ||
515 (p->flags & (PG_BUSY|PG_UNMANAGED)) ||
516 !pmap_page_exists_quick(vm_map_pmap(map), p)) {
521 actcount = pmap_ts_referenced(p);
523 vm_page_flag_set(p, PG_REFERENCED);
524 } else if (p->flags & PG_REFERENCED) {
528 if ((p->queue != PQ_ACTIVE) &&
529 (p->flags & PG_REFERENCED)) {
531 p->act_count += actcount;
532 vm_page_flag_clear(p, PG_REFERENCED);
533 } else if (p->queue == PQ_ACTIVE) {
534 if ((p->flags & PG_REFERENCED) == 0) {
535 p->act_count -= min(p->act_count, ACT_DECLINE);
536 if (!remove_mode && (vm_pageout_algorithm || (p->act_count == 0))) {
537 vm_page_protect(p, VM_PROT_NONE);
538 vm_page_deactivate(p);
540 TAILQ_REMOVE(&vm_page_queues[PQ_ACTIVE].pl, p, pageq);
541 TAILQ_INSERT_TAIL(&vm_page_queues[PQ_ACTIVE].pl, p, pageq);
545 vm_page_flag_clear(p, PG_REFERENCED);
546 if (p->act_count < (ACT_MAX - ACT_ADVANCE))
547 p->act_count += ACT_ADVANCE;
548 TAILQ_REMOVE(&vm_page_queues[PQ_ACTIVE].pl, p, pageq);
549 TAILQ_INSERT_TAIL(&vm_page_queues[PQ_ACTIVE].pl, p, pageq);
551 } else if (p->queue == PQ_INACTIVE) {
552 vm_page_protect(p, VM_PROT_NONE);
557 object = object->backing_object;
562 * deactivate some number of pages in a map, try to do it fairly, but
563 * that is really hard to do.
566 vm_pageout_map_deactivate_pages(vm_map_t map, vm_pindex_t desired)
569 vm_object_t obj, bigobj;
572 if (lockmgr(&map->lock, LK_EXCLUSIVE | LK_NOWAIT)) {
580 * first, search out the biggest object, and try to free pages from
583 tmpe = map->header.next;
584 while (tmpe != &map->header) {
585 switch(tmpe->maptype) {
586 case VM_MAPTYPE_NORMAL:
587 case VM_MAPTYPE_VPAGETABLE:
588 obj = tmpe->object.vm_object;
589 if ((obj != NULL) && (obj->shadow_count <= 1) &&
591 (bigobj->resident_page_count < obj->resident_page_count))) {
598 if (tmpe->wired_count > 0)
599 nothingwired = FALSE;
604 vm_pageout_object_deactivate_pages(map, bigobj, desired, 0);
607 * Next, hunt around for other pages to deactivate. We actually
608 * do this search sort of wrong -- .text first is not the best idea.
610 tmpe = map->header.next;
611 while (tmpe != &map->header) {
612 if (pmap_resident_count(vm_map_pmap(map)) <= desired)
614 switch(tmpe->maptype) {
615 case VM_MAPTYPE_NORMAL:
616 case VM_MAPTYPE_VPAGETABLE:
617 obj = tmpe->object.vm_object;
619 vm_pageout_object_deactivate_pages(map, obj, desired, 0);
628 * Remove all mappings if a process is swapped out, this will free page
631 if (desired == 0 && nothingwired)
632 pmap_remove(vm_map_pmap(map),
633 VM_MIN_USER_ADDRESS, VM_MAX_USER_ADDRESS);
639 * Don't try to be fancy - being fancy can lead to vnode deadlocks. We
640 * only do it for OBJT_DEFAULT and OBJT_SWAP objects which we know can
641 * be trivially freed.
644 vm_pageout_page_free(vm_page_t m) {
645 vm_object_t object = m->object;
646 int type = object->type;
648 if (type == OBJT_SWAP || type == OBJT_DEFAULT)
649 vm_object_reference(object);
651 vm_page_protect(m, VM_PROT_NONE);
653 if (type == OBJT_SWAP || type == OBJT_DEFAULT)
654 vm_object_deallocate(object);
658 * vm_pageout_scan does the dirty work for the pageout daemon.
661 struct vm_pageout_scan_info {
662 struct proc *bigproc;
666 static int vm_pageout_scan_callback(struct proc *p, void *data);
669 vm_pageout_scan(int pass)
671 struct vm_pageout_scan_info info;
673 struct vm_page marker;
674 int page_shortage, maxscan, pcount;
675 int addl_page_shortage, addl_page_shortage_init;
678 int vnodes_skipped = 0;
682 * Do whatever cleanup that the pmap code can.
686 addl_page_shortage_init = vm_pageout_deficit;
687 vm_pageout_deficit = 0;
690 * Calculate the number of pages we want to either free or move
693 page_shortage = vm_paging_target() + addl_page_shortage_init;
696 * Initialize our marker
698 bzero(&marker, sizeof(marker));
699 marker.flags = PG_BUSY | PG_FICTITIOUS | PG_MARKER;
700 marker.queue = PQ_INACTIVE;
701 marker.wire_count = 1;
704 * Start scanning the inactive queue for pages we can move to the
705 * cache or free. The scan will stop when the target is reached or
706 * we have scanned the entire inactive queue. Note that m->act_count
707 * is not used to form decisions for the inactive queue, only for the
710 * maxlaunder limits the number of dirty pages we flush per scan.
711 * For most systems a smaller value (16 or 32) is more robust under
712 * extreme memory and disk pressure because any unnecessary writes
713 * to disk can result in extreme performance degredation. However,
714 * systems with excessive dirty pages (especially when MAP_NOSYNC is
715 * used) will die horribly with limited laundering. If the pageout
716 * daemon cannot clean enough pages in the first pass, we let it go
717 * all out in succeeding passes.
719 if ((maxlaunder = vm_max_launder) <= 1)
725 * We will generally be in a critical section throughout the
726 * scan, but we can release it temporarily when we are sitting on a
727 * non-busy page without fear. this is required to prevent an
728 * interrupt from unbusying or freeing a page prior to our busy
729 * check, leaving us on the wrong queue or checking the wrong
734 addl_page_shortage = addl_page_shortage_init;
735 maxscan = vmstats.v_inactive_count;
736 for (m = TAILQ_FIRST(&vm_page_queues[PQ_INACTIVE].pl);
737 m != NULL && maxscan-- > 0 && page_shortage > 0;
740 mycpu->gd_cnt.v_pdpages++;
743 * Give interrupts a chance
749 * It's easier for some of the conditions below to just loop
750 * and catch queue changes here rather then check everywhere
753 if (m->queue != PQ_INACTIVE)
755 next = TAILQ_NEXT(m, pageq);
760 if (m->flags & PG_MARKER)
764 * A held page may be undergoing I/O, so skip it.
767 TAILQ_REMOVE(&vm_page_queues[PQ_INACTIVE].pl, m, pageq);
768 TAILQ_INSERT_TAIL(&vm_page_queues[PQ_INACTIVE].pl, m, pageq);
769 addl_page_shortage++;
774 * Dont mess with busy pages, keep in the front of the
775 * queue, most likely are being paged out.
777 if (m->busy || (m->flags & PG_BUSY)) {
778 addl_page_shortage++;
782 if (m->object->ref_count == 0) {
784 * If the object is not being used, we ignore previous
787 vm_page_flag_clear(m, PG_REFERENCED);
788 pmap_clear_reference(m);
790 } else if (((m->flags & PG_REFERENCED) == 0) &&
791 (actcount = pmap_ts_referenced(m))) {
793 * Otherwise, if the page has been referenced while
794 * in the inactive queue, we bump the "activation
795 * count" upwards, making it less likely that the
796 * page will be added back to the inactive queue
797 * prematurely again. Here we check the page tables
798 * (or emulated bits, if any), given the upper level
799 * VM system not knowing anything about existing
803 m->act_count += (actcount + ACT_ADVANCE);
808 * If the upper level VM system knows about any page
809 * references, we activate the page. We also set the
810 * "activation count" higher than normal so that we will less
811 * likely place pages back onto the inactive queue again.
813 if ((m->flags & PG_REFERENCED) != 0) {
814 vm_page_flag_clear(m, PG_REFERENCED);
815 actcount = pmap_ts_referenced(m);
817 m->act_count += (actcount + ACT_ADVANCE + 1);
822 * If the upper level VM system doesn't know anything about
823 * the page being dirty, we have to check for it again. As
824 * far as the VM code knows, any partially dirty pages are
827 * Pages marked PG_WRITEABLE may be mapped into the user
828 * address space of a process running on another cpu. A
829 * user process (without holding the MP lock) running on
830 * another cpu may be able to touch the page while we are
831 * trying to remove it. To prevent this from occuring we
832 * must call pmap_remove_all() or otherwise make the page
833 * read-only. If the race occured pmap_remove_all() is
834 * responsible for setting m->dirty.
837 vm_page_test_dirty(m);
839 if (m->dirty == 0 && (m->flags & PG_WRITEABLE) != 0)
848 * Invalid pages can be easily freed
850 vm_pageout_page_free(m);
851 mycpu->gd_cnt.v_dfree++;
853 } else if (m->dirty == 0) {
855 * Clean pages can be placed onto the cache queue.
856 * This effectively frees them.
860 } else if ((m->flags & PG_WINATCFLS) == 0 && pass == 0) {
862 * Dirty pages need to be paged out, but flushing
863 * a page is extremely expensive verses freeing
864 * a clean page. Rather then artificially limiting
865 * the number of pages we can flush, we instead give
866 * dirty pages extra priority on the inactive queue
867 * by forcing them to be cycled through the queue
868 * twice before being flushed, after which the
869 * (now clean) page will cycle through once more
870 * before being freed. This significantly extends
871 * the thrash point for a heavily loaded machine.
873 vm_page_flag_set(m, PG_WINATCFLS);
874 TAILQ_REMOVE(&vm_page_queues[PQ_INACTIVE].pl, m, pageq);
875 TAILQ_INSERT_TAIL(&vm_page_queues[PQ_INACTIVE].pl, m, pageq);
876 } else if (maxlaunder > 0) {
878 * We always want to try to flush some dirty pages if
879 * we encounter them, to keep the system stable.
880 * Normally this number is small, but under extreme
881 * pressure where there are insufficient clean pages
882 * on the inactive queue, we may have to go all out.
884 int swap_pageouts_ok;
885 struct vnode *vp = NULL;
889 if ((object->type != OBJT_SWAP) && (object->type != OBJT_DEFAULT)) {
890 swap_pageouts_ok = 1;
892 swap_pageouts_ok = !(defer_swap_pageouts || disable_swap_pageouts);
893 swap_pageouts_ok |= (!disable_swap_pageouts && defer_swap_pageouts &&
894 vm_page_count_min());
899 * We don't bother paging objects that are "dead".
900 * Those objects are in a "rundown" state.
902 if (!swap_pageouts_ok || (object->flags & OBJ_DEAD)) {
903 TAILQ_REMOVE(&vm_page_queues[PQ_INACTIVE].pl, m, pageq);
904 TAILQ_INSERT_TAIL(&vm_page_queues[PQ_INACTIVE].pl, m, pageq);
909 * The object is already known NOT to be dead. It
910 * is possible for the vget() to block the whole
911 * pageout daemon, but the new low-memory handling
912 * code should prevent it.
914 * The previous code skipped locked vnodes and, worse,
915 * reordered pages in the queue. This results in
916 * completely non-deterministic operation because,
917 * quite often, a vm_fault has initiated an I/O and
918 * is holding a locked vnode at just the point where
919 * the pageout daemon is woken up.
921 * We can't wait forever for the vnode lock, we might
922 * deadlock due to a vn_read() getting stuck in
923 * vm_wait while holding this vnode. We skip the
924 * vnode if we can't get it in a reasonable amount
928 if (object->type == OBJT_VNODE) {
931 if (vget(vp, LK_EXCLUSIVE|LK_NOOBJ|LK_TIMELOCK)) {
933 if (object->flags & OBJ_MIGHTBEDIRTY)
939 * The page might have been moved to another
940 * queue during potential blocking in vget()
941 * above. The page might have been freed and
942 * reused for another vnode. The object might
943 * have been reused for another vnode.
945 if (m->queue != PQ_INACTIVE ||
946 m->object != object ||
947 object->handle != vp) {
948 if (object->flags & OBJ_MIGHTBEDIRTY)
955 * The page may have been busied during the
956 * blocking in vput(); We don't move the
957 * page back onto the end of the queue so that
958 * statistics are more correct if we don't.
960 if (m->busy || (m->flags & PG_BUSY)) {
966 * If the page has become held it might
967 * be undergoing I/O, so skip it
970 TAILQ_REMOVE(&vm_page_queues[PQ_INACTIVE].pl, m, pageq);
971 TAILQ_INSERT_TAIL(&vm_page_queues[PQ_INACTIVE].pl, m, pageq);
972 if (object->flags & OBJ_MIGHTBEDIRTY)
980 * If a page is dirty, then it is either being washed
981 * (but not yet cleaned) or it is still in the
982 * laundry. If it is still in the laundry, then we
983 * start the cleaning operation.
985 * This operation may cluster, invalidating the 'next'
986 * pointer. To prevent an inordinate number of
987 * restarts we use our marker to remember our place.
989 * decrement page_shortage on success to account for
990 * the (future) cleaned page. Otherwise we could wind
991 * up laundering or cleaning too many pages.
993 TAILQ_INSERT_AFTER(&vm_page_queues[PQ_INACTIVE].pl, m, &marker, pageq);
994 if (vm_pageout_clean(m) != 0) {
998 next = TAILQ_NEXT(&marker, pageq);
999 TAILQ_REMOVE(&vm_page_queues[PQ_INACTIVE].pl, &marker, pageq);
1006 * Compute the number of pages we want to try to move from the
1007 * active queue to the inactive queue.
1009 page_shortage = vm_paging_target() +
1010 vmstats.v_inactive_target - vmstats.v_inactive_count;
1011 page_shortage += addl_page_shortage;
1014 * Scan the active queue for things we can deactivate. We nominally
1015 * track the per-page activity counter and use it to locate
1016 * deactivation candidates.
1018 * NOTE: we are still in a critical section.
1020 pcount = vmstats.v_active_count;
1021 m = TAILQ_FIRST(&vm_page_queues[PQ_ACTIVE].pl);
1023 while ((m != NULL) && (pcount-- > 0) && (page_shortage > 0)) {
1025 * Give interrupts a chance.
1031 * If the page was ripped out from under us, just stop.
1033 if (m->queue != PQ_ACTIVE)
1035 next = TAILQ_NEXT(m, pageq);
1038 * Don't deactivate pages that are busy.
1040 if ((m->busy != 0) ||
1041 (m->flags & PG_BUSY) ||
1042 (m->hold_count != 0)) {
1043 TAILQ_REMOVE(&vm_page_queues[PQ_ACTIVE].pl, m, pageq);
1044 TAILQ_INSERT_TAIL(&vm_page_queues[PQ_ACTIVE].pl, m, pageq);
1050 * The count for pagedaemon pages is done after checking the
1051 * page for eligibility...
1053 mycpu->gd_cnt.v_pdpages++;
1056 * Check to see "how much" the page has been used.
1059 if (m->object->ref_count != 0) {
1060 if (m->flags & PG_REFERENCED) {
1063 actcount += pmap_ts_referenced(m);
1065 m->act_count += ACT_ADVANCE + actcount;
1066 if (m->act_count > ACT_MAX)
1067 m->act_count = ACT_MAX;
1072 * Since we have "tested" this bit, we need to clear it now.
1074 vm_page_flag_clear(m, PG_REFERENCED);
1077 * Only if an object is currently being used, do we use the
1078 * page activation count stats.
1080 if (actcount && (m->object->ref_count != 0)) {
1081 TAILQ_REMOVE(&vm_page_queues[PQ_ACTIVE].pl, m, pageq);
1082 TAILQ_INSERT_TAIL(&vm_page_queues[PQ_ACTIVE].pl, m, pageq);
1084 m->act_count -= min(m->act_count, ACT_DECLINE);
1085 if (vm_pageout_algorithm ||
1086 m->object->ref_count == 0 ||
1087 m->act_count < pass) {
1089 if (m->object->ref_count == 0) {
1090 vm_page_protect(m, VM_PROT_NONE);
1094 vm_page_deactivate(m);
1096 vm_page_deactivate(m);
1099 TAILQ_REMOVE(&vm_page_queues[PQ_ACTIVE].pl, m, pageq);
1100 TAILQ_INSERT_TAIL(&vm_page_queues[PQ_ACTIVE].pl, m, pageq);
1107 * We try to maintain some *really* free pages, this allows interrupt
1108 * code to be guaranteed space. Since both cache and free queues
1109 * are considered basically 'free', moving pages from cache to free
1110 * does not effect other calculations.
1112 * NOTE: we are still in a critical section.
1115 while (vmstats.v_free_count < vmstats.v_free_reserved) {
1116 static int cache_rover = 0;
1117 m = vm_page_list_find(PQ_CACHE, cache_rover, FALSE);
1120 if ((m->flags & (PG_BUSY|PG_UNMANAGED)) ||
1125 printf("Warning: busy page %p found in cache\n", m);
1127 vm_page_deactivate(m);
1130 cache_rover = (cache_rover + PQ_PRIME2) & PQ_L2_MASK;
1131 vm_pageout_page_free(m);
1132 mycpu->gd_cnt.v_dfree++;
1137 #if !defined(NO_SWAPPING)
1139 * Idle process swapout -- run once per second.
1141 if (vm_swap_idle_enabled) {
1143 if (time_second != lsec) {
1144 vm_pageout_req_swapout |= VM_SWAP_IDLE;
1152 * If we didn't get enough free pages, and we have skipped a vnode
1153 * in a writeable object, wakeup the sync daemon. And kick swapout
1154 * if we did not get enough free pages.
1156 if (vm_paging_target() > 0) {
1157 if (vnodes_skipped && vm_page_count_min())
1159 #if !defined(NO_SWAPPING)
1160 if (vm_swap_enabled && vm_page_count_target()) {
1162 vm_pageout_req_swapout |= VM_SWAP_NORMAL;
1168 * If we are out of swap and were not able to reach our paging
1169 * target, kill the largest process.
1171 if ((vm_swap_size < 64 && vm_page_count_min()) ||
1172 (swap_pager_full && vm_paging_target() > 0)) {
1174 if ((vm_swap_size < 64 || swap_pager_full) && vm_page_count_min()) {
1176 info.bigproc = NULL;
1178 allproc_scan(vm_pageout_scan_callback, &info);
1179 if (info.bigproc != NULL) {
1180 killproc(info.bigproc, "out of swap space");
1181 info.bigproc->p_nice = PRIO_MIN;
1182 info.bigproc->p_usched->resetpriority(&info.bigproc->p_lwp);
1183 wakeup(&vmstats.v_free_count);
1184 PRELE(info.bigproc);
1190 vm_pageout_scan_callback(struct proc *p, void *data)
1192 struct vm_pageout_scan_info *info = data;
1196 * if this is a system process, skip it
1198 if ((p->p_flag & P_SYSTEM) || (p->p_pid == 1) ||
1199 ((p->p_pid < 48) && (vm_swap_size != 0))) {
1204 * if the process is in a non-running type state,
1207 if (p->p_stat != SRUN && p->p_stat != SSLEEP) {
1212 * get the process size
1214 size = vmspace_resident_count(p->p_vmspace) +
1215 vmspace_swap_count(p->p_vmspace);
1218 * If the this process is bigger than the biggest one
1221 if (size > info->bigsize) {
1223 PRELE(info->bigproc);
1226 info->bigsize = size;
1232 * This routine tries to maintain the pseudo LRU active queue,
1233 * so that during long periods of time where there is no paging,
1234 * that some statistic accumulation still occurs. This code
1235 * helps the situation where paging just starts to occur.
1238 vm_pageout_page_stats(void)
1241 int pcount,tpcount; /* Number of pages to check */
1242 static int fullintervalcount = 0;
1246 (vmstats.v_inactive_target + vmstats.v_cache_max + vmstats.v_free_min) -
1247 (vmstats.v_free_count + vmstats.v_inactive_count + vmstats.v_cache_count);
1249 if (page_shortage <= 0)
1254 pcount = vmstats.v_active_count;
1255 fullintervalcount += vm_pageout_stats_interval;
1256 if (fullintervalcount < vm_pageout_full_stats_interval) {
1257 tpcount = (vm_pageout_stats_max * vmstats.v_active_count) / vmstats.v_page_count;
1258 if (pcount > tpcount)
1261 fullintervalcount = 0;
1264 m = TAILQ_FIRST(&vm_page_queues[PQ_ACTIVE].pl);
1265 while ((m != NULL) && (pcount-- > 0)) {
1268 if (m->queue != PQ_ACTIVE) {
1272 next = TAILQ_NEXT(m, pageq);
1274 * Don't deactivate pages that are busy.
1276 if ((m->busy != 0) ||
1277 (m->flags & PG_BUSY) ||
1278 (m->hold_count != 0)) {
1279 TAILQ_REMOVE(&vm_page_queues[PQ_ACTIVE].pl, m, pageq);
1280 TAILQ_INSERT_TAIL(&vm_page_queues[PQ_ACTIVE].pl, m, pageq);
1286 if (m->flags & PG_REFERENCED) {
1287 vm_page_flag_clear(m, PG_REFERENCED);
1291 actcount += pmap_ts_referenced(m);
1293 m->act_count += ACT_ADVANCE + actcount;
1294 if (m->act_count > ACT_MAX)
1295 m->act_count = ACT_MAX;
1296 TAILQ_REMOVE(&vm_page_queues[PQ_ACTIVE].pl, m, pageq);
1297 TAILQ_INSERT_TAIL(&vm_page_queues[PQ_ACTIVE].pl, m, pageq);
1299 if (m->act_count == 0) {
1301 * We turn off page access, so that we have
1302 * more accurate RSS stats. We don't do this
1303 * in the normal page deactivation when the
1304 * system is loaded VM wise, because the
1305 * cost of the large number of page protect
1306 * operations would be higher than the value
1307 * of doing the operation.
1309 vm_page_protect(m, VM_PROT_NONE);
1310 vm_page_deactivate(m);
1312 m->act_count -= min(m->act_count, ACT_DECLINE);
1313 TAILQ_REMOVE(&vm_page_queues[PQ_ACTIVE].pl, m, pageq);
1314 TAILQ_INSERT_TAIL(&vm_page_queues[PQ_ACTIVE].pl, m, pageq);
1324 vm_pageout_free_page_calc(vm_size_t count)
1326 if (count < vmstats.v_page_count)
1329 * free_reserved needs to include enough for the largest swap pager
1330 * structures plus enough for any pv_entry structs when paging.
1332 if (vmstats.v_page_count > 1024)
1333 vmstats.v_free_min = 4 + (vmstats.v_page_count - 1024) / 200;
1335 vmstats.v_free_min = 4;
1336 vmstats.v_pageout_free_min = (2*MAXBSIZE)/PAGE_SIZE +
1337 vmstats.v_interrupt_free_min;
1338 vmstats.v_free_reserved = vm_pageout_page_count +
1339 vmstats.v_pageout_free_min + (count / 768) + PQ_L2_SIZE;
1340 vmstats.v_free_severe = vmstats.v_free_min / 2;
1341 vmstats.v_free_min += vmstats.v_free_reserved;
1342 vmstats.v_free_severe += vmstats.v_free_reserved;
1348 * vm_pageout is the high level pageout daemon.
1356 * Initialize some paging parameters.
1359 vmstats.v_interrupt_free_min = 2;
1360 if (vmstats.v_page_count < 2000)
1361 vm_pageout_page_count = 8;
1363 vm_pageout_free_page_calc(vmstats.v_page_count);
1365 * v_free_target and v_cache_min control pageout hysteresis. Note
1366 * that these are more a measure of the VM cache queue hysteresis
1367 * then the VM free queue. Specifically, v_free_target is the
1368 * high water mark (free+cache pages).
1370 * v_free_reserved + v_cache_min (mostly means v_cache_min) is the
1371 * low water mark, while v_free_min is the stop. v_cache_min must
1372 * be big enough to handle memory needs while the pageout daemon
1373 * is signalled and run to free more pages.
1375 if (vmstats.v_free_count > 6144)
1376 vmstats.v_free_target = 4 * vmstats.v_free_min + vmstats.v_free_reserved;
1378 vmstats.v_free_target = 2 * vmstats.v_free_min + vmstats.v_free_reserved;
1380 if (vmstats.v_free_count > 2048) {
1381 vmstats.v_cache_min = vmstats.v_free_target;
1382 vmstats.v_cache_max = 2 * vmstats.v_cache_min;
1383 vmstats.v_inactive_target = (3 * vmstats.v_free_target) / 2;
1385 vmstats.v_cache_min = 0;
1386 vmstats.v_cache_max = 0;
1387 vmstats.v_inactive_target = vmstats.v_free_count / 4;
1389 if (vmstats.v_inactive_target > vmstats.v_free_count / 3)
1390 vmstats.v_inactive_target = vmstats.v_free_count / 3;
1392 /* XXX does not really belong here */
1393 if (vm_page_max_wired == 0)
1394 vm_page_max_wired = vmstats.v_free_count / 3;
1396 if (vm_pageout_stats_max == 0)
1397 vm_pageout_stats_max = vmstats.v_free_target;
1400 * Set interval in seconds for stats scan.
1402 if (vm_pageout_stats_interval == 0)
1403 vm_pageout_stats_interval = 5;
1404 if (vm_pageout_full_stats_interval == 0)
1405 vm_pageout_full_stats_interval = vm_pageout_stats_interval * 4;
1409 * Set maximum free per pass
1411 if (vm_pageout_stats_free_max == 0)
1412 vm_pageout_stats_free_max = 5;
1414 swap_pager_swap_init();
1417 * The pageout daemon is never done, so loop forever.
1423 * If we have enough free memory, wakeup waiters. Do
1424 * not clear vm_pages_needed until we reach our target,
1425 * otherwise we may be woken up over and over again and
1426 * waste a lot of cpu.
1429 if (vm_pages_needed && !vm_page_count_min()) {
1430 if (vm_paging_needed() <= 0)
1431 vm_pages_needed = 0;
1432 wakeup(&vmstats.v_free_count);
1434 if (vm_pages_needed) {
1436 * Still not done, take a second pass without waiting
1437 * (unlimited dirty cleaning), otherwise sleep a bit
1442 tsleep(&vm_pages_needed, 0, "psleep", hz/2);
1445 * Good enough, sleep & handle stats. Prime the pass
1452 error = tsleep(&vm_pages_needed,
1453 0, "psleep", vm_pageout_stats_interval * hz);
1454 if (error && !vm_pages_needed) {
1457 vm_pageout_page_stats();
1462 if (vm_pages_needed)
1463 mycpu->gd_cnt.v_pdwakeups++;
1465 vm_pageout_scan(pass);
1466 vm_pageout_deficit = 0;
1471 pagedaemon_wakeup(void)
1473 if (!vm_pages_needed && curthread != pagethread) {
1475 wakeup(&vm_pages_needed);
1479 #if !defined(NO_SWAPPING)
1481 vm_req_vmdaemon(void)
1483 static int lastrun = 0;
1485 if ((ticks > (lastrun + hz)) || (ticks < lastrun)) {
1486 wakeup(&vm_daemon_needed);
1491 static int vm_daemon_callback(struct proc *p, void *data __unused);
1497 tsleep(&vm_daemon_needed, 0, "psleep", 0);
1498 if (vm_pageout_req_swapout) {
1499 swapout_procs(vm_pageout_req_swapout);
1500 vm_pageout_req_swapout = 0;
1503 * scan the processes for exceeding their rlimits or if
1504 * process is swapped out -- deactivate pages
1506 allproc_scan(vm_daemon_callback, NULL);
1511 vm_daemon_callback(struct proc *p, void *data __unused)
1513 vm_pindex_t limit, size;
1516 * if this is a system process or if we have already
1517 * looked at this process, skip it.
1519 if (p->p_flag & (P_SYSTEM | P_WEXIT))
1523 * if the process is in a non-running type state,
1526 if (p->p_stat != SRUN && p->p_stat != SSLEEP)
1532 limit = OFF_TO_IDX(qmin(p->p_rlimit[RLIMIT_RSS].rlim_cur,
1533 p->p_rlimit[RLIMIT_RSS].rlim_max));
1536 * let processes that are swapped out really be
1537 * swapped out. Set the limit to nothing to get as
1538 * many pages out to swap as possible.
1540 if (p->p_flag & P_SWAPPEDOUT)
1543 size = vmspace_resident_count(p->p_vmspace);
1544 if (limit >= 0 && size >= limit) {
1545 vm_pageout_map_deactivate_pages(
1546 &p->p_vmspace->vm_map, limit);